sgl-project/sglang · critical · ValueError

Kimi expert-pack identity mismatch at index {index}

Error message

Kimi expert-pack identity mismatch at index {index}

What it means

Index entry fields don't match the derived identity: the entry name must parse as layer/role, expert must equal the derived expected expert for the position, and reserved must be 0. This is a structural integrity check that entries appear in the exact canonical order.

Source

Thrown at python/sglang/srt/layers/moe/expert_pack.py:835

                    raise ValueError("Kimi expert-pack index is truncated")
                index_digest.update(raw_entry)
                name_raw, expert, reserved, offset, nbytes = GGML_PACK_ENTRY.unpack(
                    raw_entry
                )
                object_index, physical_role_id = divmod(index, len(KIMI_PHYSICAL_ROLES))
                layer_index, expected_expert = divmod(object_index, expected_experts)
                layer = active_layers[layer_index]
                physical_role = KIMI_PHYSICAL_ROLES[physical_role_id]
                name = _fixed_string(name_raw)
                match = KIMI_EXPERT_RE.fullmatch(name)
                if (
                    match is None
                    or int(match.group("layer")) != layer
                    or match.group("role") != physical_role
                    or expert != expected_expert
                    or reserved != 0
                ):
                    raise ValueError(
                        f"Kimi expert-pack identity mismatch at index {index}"
                    )
                expected_nbytes = role_nbytes[physical_role]
                if (
                    nbytes != expected_nbytes
                    or offset % int(pack_manifest["alignment"])
                    or offset < previous_end
                    or offset + nbytes > self.path.stat().st_size
                ):
                    raise ValueError(
                        f"Kimi expert-pack range mismatch at index {index}"
                    )
                previous_end = offset + nbytes
                object_key = (layer, expert)
                if physical_role_id == 0:
                    self.object_offsets[object_key] = offset
                expected_offset = (
                    self.object_offsets[object_key] + role_offsets[physical_role]

View on GitHub (pinned to 0132848349)

Solutions

1. Re-pack with the official tool so entries are written in canonical layer-major, role-minor order with correct names
2. If writing your own packer, emit name f-strings matching layer/role, expert = index-derived value, reserved = 0

try:
    pack = KimiExpertPack(path, manifest)
except ValueError as e:
    if 'identity mismatch' in str(e):
        log.error('pack index entries out of canonical order; re-pack needed')
    raise

When it happens

Trigger: Entries shuffled or renamed (non-canonical ordering), a reserved field set nonzero by a foreign writer, or a pack from a tool that names entries differently.

Common situations: See trigger scenarios.

Related errors

• Kimi active routed MoE layers must be exactly 1..92
• Kimi manifest expert-pack path does not match pack_path
• Kimi expert-pack size does not match its manifest
• Kimi expert-pack physical role order is unsupported
• Kimi expert-pack role sizes do not match object bytes

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/7b5fb3f72b016592. Report an issue: GitHub.